Dataset Card for Finance Commons AMF OCR dataset (FC-AMF-OCR)

Dataset Summary

The FC-AMF-OCR dataset is a comprehensive document collection derived from the AMF-PDF dataset, which is part of the Finance Commons collection. This extensive dataset comprises 9.3 million images, each processed through Optical Character Recognition (OCR) using the docTR library. While native text annotations are available in the AMF-Text dataset, these annotations suffer from imperfections and inaccuracies, including mainly missing spaces, extra spaces, artifacts, etc. Additionally, the format of these annotations — presented as a single, continuous block of text without page demarcations — limits their utility for image-to-text tasks.

The FC-AMF-OCR dataset aims to address these limitations by providing:

Full bounding box information for each element
Confidence scores for individual words, lines, and text blocks
Per-page annotations instead of a single block of text per document
Solve the space inaccuracies in the native text annotations

An example page of one pdf document with existing text annotation(red) and the OCR annotation(green). For simplicity, we order text from left to right and top to bottom.

Most existing large scale OCR datasets like the Industry Documents Library (IDL) or the PDF Association dataset (PDFA) suffer from a number of issues:

Time Coverage: These datasets consist primarily of older documents or PDFs from specific periods, which might not reflect current trends or developments.
OCR Engines: They use outdated or inconsistent OCR technologies, affecting the accuracy and reliability of text extraction.
Further, some of these annotations are limited to what can be extracted and is readily available - text drawn in images and only present as bitmap renditions is missed entirely.

FC-AMF-OCR enhances existing datasets by offering detailed OCR annotations for a recent collection of text-rich documents from the French Authority for Financial Markets (AMF). It leverages the excellent open-source docTR OCR engine to extract text from various elements, including images and logos. By utilizing an open-source solution, FC-AMF-OCR ensures stability against API changes and allows users to implement custom filtering as needed. This approach provides researchers and developers with a reliable and transparent tool for comprehensive document understanding and analysis.

Following most large scale OCR datasets like IDL, this dataset is also in webdataset .tar format and can be used with the webdataset library in a seamless way. Concretely, each document exists as a pair of a pdf and a json.gz file containing the OCR annotation.

Load the dataset with `datasets`

This dataset can be used with Hugging Face datasets. Here is an example of how to stream the dataset directly from Hugging Face so you don't have to download the dataset locally.

Note: We do recommend downloading the dataset to speed up the processing.

from datasets import load_dataset

dataset = load_dataset('lightonai/fc-amf-ocr', streaming=True)
print(next(iter(dataset['train'])).keys())
>> dict_keys(['__key__', '__url__', 'pdf', 'json.gz'])

You can download the dataset using the following command:

 import os
 from huggingface_hub import HfApi
 
 os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
 api = HfApi()
 api.snapshot_download("lightonai/fc-amf-ocr", repo_type="dataset", local_dir_use_symlinks=False)

Approach

We start from the original dataset, which is a collection of 633,244 PDF files and apply some simple filters to remove files that are not relevant for training. The main goal is to have a dataset that is ready to use for large-scale training. We use the following filters:

Corrupted files: we remove files that fail to be decoded correctly or that take too long to load.
Page count: we remove files that have more than 500 pages. Large files take too long to load and render.
Keep original quality: we apply no compression or rendering that would degrade the quality of the original PDF.

The basic filtering removes less than 1% of the original dataset. After the basic filtering:

We selected the best performing models from the docTR library. For maximum accuracy, we keep all models in full precision(FP32).

detection model : DBNet with a ResNet-50 backbone
recognition model : CRNN with a VGG-16 backbone

We use data-parallel to parallelize the OCR process over multiple GPUs. This is done by splitting the dataset into multiple shards and processing each shard in parallel.
The recognition model is compiled with torch.compile to speed up the inference.

By default the images are rendered at a DPI of 144 for all the processing steps but we provide the original PDFs so users can render them at their preffered quality. Having access to the full PDF quality is very important for training robust models.

The dataset's page distribution is represented in the following histogram. On average, documents contain approximately 15 pages, while the median page count is about 2.

The distribution of number of pages in the FC-AMF-OCR dataset.

We also show the year distribution of the dataset. The dataset contains documents from 2008 to 2024. This shows that the dataset is relatively recent and covers a wide range of years which complements previous datasets.

The distribution of years in the FC-AMF-OCR dataset.

How to visualize a page from the dataset?

PDF files are sourced from a variety of origins and are typically stored in RGB format. These files can consist of multiple pages, each of which can be rendered using different tools or engines according to your needs. One recommended option is pdf2image, a tool that converts PDF pages into images. To use pdf2image, you need to install the poppler-utils package, which provides the necessary support for rendering and processing PDF files efficiently. This approach allows for flexible handling of PDFs, making it easier to extract and manipulate content from multi-page documents.

apt-get install poppler-utils

from pdf2image import convert_from_bytes

page = convert_from_bytes(sample['pdf'], dpi=300, first_page=1, last_page=1)[0]
page

A page from the FC-AMF-OCR dataset.

Each pdf is paired with a json.gz file with the structure shown below. This strucure is that of docTR outputs, you can learn more here: here. We explicitly avoid applying any OCR post-processing to get an approximate reading order. There are multiple ways of getting a reading order from bounding boxes. Users can use their own heuristics to extract the reading order from the bounding boxes.

{
    'pages': [{
        'page_idx': 0,
        'dimensions': [1684, 1191],
        'geometry': [[0.2514, 0.0712], [0.4144, 0.0986]],
        'lines': [{
            'geometry': [[0.2515, 0.0713], [0.4144, 0.0850]],
            'words': [
                {
                    'value': 'European',
                    'confidence': 0.9998,
                    'geometry': [[0.2515, 0.0732], [0.3150, 0.0850]]
                },
                {
                    'value': 'Securities',
                    'confidence': 0.9986,
                    'geometry': [[0.3164, 0.0723], [0.3840, 0.0830]]
                },
                {
                    'value': 'and',
                    'confidence': 0.9998,
                    'geometry': [[0.3854, 0.0713], [0.4144, 0.0840]]
                }
            ]
        },
        {
            'geometry': [[0.2515, 0.0840], [0.3730, 0.0986]],
            'words': [
                {
                    'value': 'Markets',
                    'confidence': 0.9977,
                    'geometry': [[0.2515, 0.0850], [0.3053, 0.0967]]
                },
                {
                    'value': 'Authority',
                    'confidence': 0.8129,
                    'geometry': [[0.3067, 0.0840], [0.3730, 0.0986]]
                }
            ]
        }]
    }]
}

Document Structure

The structural organization of the documents, including words, lines, blocks, pages, and the overall document is as follows.

Element	Description
Word	A Word is an uninterrupted sequence of characters.
Line	A collection of Words aligned spatially and meant to be read together.
Block	A collection of Lines.
Page	A collection of Blocks that were on the same physical page.

The top-level key, pages, is a list containing each page in the document. In this example, only one page is shown.

Page:
- page_idx: The index of the page in the document (starts at 0).
- dimensions: The dimensions of the page in pixels, formatted as [height, width].
Blocks:
- A page consists of several blocks, each containing lines.
- geometry: Defines the bounding box of the block using normalized coordinates relative to the page size.
Lines:
- Each block contains a list of lines, where a line is a sequence of words grouped together.
- geometry: Bounding box of the line in normalized coordinates relative to the page size.
Words:
- Each line is composed of individual words (continuous sequences of characters).
- value: The text content of the word.
- confidence: The confidence score of the OCR engine for the word.
- geometry: Bounding box of the word in normalized coordinates relative to the page size.

For each page, the structure includes:

Blocks: Grouped lines within a page.
Lines: Sequences of words within a block.
Words: Individual characters or words detected within each line, along with their confidence scores and positions.

Bounding box visualization

You can visualize the bounding boxes of the dataset using the following code snippet. This code uses the pdf2image library to convert the PDF files to images.

import gzip

import matplotlib.pyplot as plt
import matplotlib.patches as patches
from matplotlib.collections import PatchCollection
from pdf2image import convert_from_path


def visualize_bounding_boxes(pdf_path, json_path, page_num=0):
    with gzip.open(json_path, 'rt', encoding='utf-8') as f:
        json_data = json.load(f)

    image = convert_from_path(pdf_path)[page_num]
    img_width, img_height = image.size

    fig, ax = plt.subplots(1, figsize=(20, 20))
    ax.imshow(image)

    patches_list = []

    for block in json_data['pages'][page_num]['blocks']:
        for line in block['lines']:
            for word in line['words']:
                bbox = word['geometry']
                x1, y1 = bbox[0]
                x2, y2 = bbox[1]

                x1, y1 = x1 * img_width, y1 * img_height
                x2, y2 = x2 * img_width, y2 * img_height

                width = x2 - x1
                height = y2 - y1

                rect = patches.Rectangle((x1, y1), width, height, linewidth=1, edgecolor='r', facecolor='none')
                patches_list.append(rect)

    patch_collection = PatchCollection(patches_list, match_original=True)
    ax.add_collection(patch_collection)

    plt.axis('off')
    plt.tight_layout()
    plt.show()

Visualizing all bounding boxes on a given page, we obtain the following:

An example page with bounding box annotations in the FC-AMF-OCR dataset.

Data Splits

There is only a single train split for this dataset.

Train

fc-amf-train-{0000..0838}.tar
838 shards (each shard is around 500 MB)
605,438 PDF files or samples
9.3M pages

Additional Information

Compute

The compute was carried out on an HPE Cray node with 8xH100, hosted on Orange Business Cloud Avenue.

Note

This dataset is intended as an OCR-heavy pre-training task for vision-language models or specialized OCR models. The current version contains multilingual data with English and French as the most represented languages. The OCR annotation might not work well for other languages due to the OCR engine limitations. Filtering based on word confidence scores can be used as a heuristic to subsample the dataset for higher quality. This approach can be scaled further by using a larger dataset with more languages and more diverse content, making it a reliable way to get multimodal data for documents.

Licensing Information

Data has been OCRed from the original dataset. As a consequence it has the same AMF-PDF license.

Note: This dataset card template was inspired by the PDFA/IDL dataset cards.